Academic Journal
Domain generalisation challenges in breast cancer molecular classification using foundation models: a cross-cohort exploratory study.
| Τίτλος: | Domain generalisation challenges in breast cancer molecular classification using foundation models: a cross-cohort exploratory study. |
|---|---|
| Συγγραφείς: | Fernandez-Romero J; Department of Computer Languages and Systems, ETSII, University of Seville, Av. Reina Mercedes s/n, 41012, Seville, Andalusia, Spain., Ramos-Berciano P; Department of Computer Languages and Systems, ETSII, University of Seville, Av. Reina Mercedes s/n, 41012, Seville, Andalusia, Spain., Perez-Perez M; Department of Normal and Pathological Cytology and Histology, Faculty of Medicine, University of Seville, Av. Doctor Fadriani s/n, 41009, Seville, Andalusia, Spain., Benavides D; Department of Computer Languages and Systems, ETSII, University of Seville, Av. Reina Mercedes s/n, 41012, Seville, Andalusia, Spain., Robles-Frias A; UGC Pathology, Hospital Universitario Virgen de Valme, Ctra. de Cádiz Km. 548, 41004, Seville, Andalusia, Spain., Garcia-Gutierrez J; Department of Computer Languages and Systems, ETSII, University of Seville, Av. Reina Mercedes s/n, 41012, Seville, Andalusia, Spain. jorgarcia@us.es., Macias-Garcia L; Department of Normal and Pathological Cytology and Histology, Faculty of Medicine, University of Seville, Av. Doctor Fadriani s/n, 41009, Seville, Andalusia, Spain. |
| Πηγή: | Medical & biological engineering & computing [Med Biol Eng Comput] 2026 Jun; Vol. 64 (6), pp. 2321-2331. Date of Electronic Publication: 2026 May 11. |
| Τύπος έκδοσης: | Journal Article |
| Γλώσσα: | English |
| Στοιχεία περιοδικού: | Publisher: Springer Country of Publication: United States NLM ID: 7704869 Publication Model: Print-Electronic Cited Medium: Internet ISSN: 1741-0444 (Electronic) Linking ISSN: 01400118 NLM ISO Abbreviation: Med Biol Eng Comput Subsets: MEDLINE |
| Imprint Name(s): | Publication: New York, NY : Springer Original Publication: Stevenage, Eng., Peregrinus. |
| Ιατρικοί όροι (MeSH): | Breast Neoplasms*/classification , Breast Neoplasms*/genetics , Breast Neoplasms*/metabolism , Breast Neoplasms*/pathology , Classification Algorithms* , Multiple-Instance Learning Algorithms*, Biomarkers, Tumor/metabolism ; Erb-b2 Receptor Tyrosine Kinases/metabolism ; Female ; Humans ; Cohort Studies ; Immunohistochemistry |
| Περίληψη: | Molecular classification guides breast cancer treatment, but PAM50 and immunohistochemistry (IHC) remain costly and unavailable in many settings. Foundation models (FMs) combined with multiple instance learning (MIL) show promise for predicting molecular subtypes from haematoxylin-and-eosin-stained slides, yet most studies report only internal validation. This study evaluates FMs with MIL across cohorts and identifies factors associated with domain-induced performance degradation. We evaluate 13 FMs and 3 complementary MIL architectures for PAM50 subtyping and IHC biomarker prediction using cross-validation on TCGA-BRCA ([Formula: see text]) and external validation on CPTAC-BRCA ([Formula: see text]). Virchow v2 achieves the best overall performance but exhibits severe degradation upon external validation, consistent across all three MIL architectures especially for HER2-enriched and Normal-like PAM50 subtypes and HER2-positive IHC prediction. Four hypothesised domain shift factors are quantified through exploratory regression analysis to explain relative performance drop (RPD). Staining variability, feature space divergence and morphological separability reach significance in univariate analysis, whilst prevalence shift does not. Staining variability and feature space divergence as covariate-level factors jointly account for 80.0% of RPD variance in the most parsimonious multivariate model ([Formula: see text], [Formula: see text]). Although based on a limited number of class-level observations and therefore exploratory in nature, these findings highlight the need for domain generalisation strategies targeting covariate shift, even when specialised FMs are used as feature encoders. (© 2026. The Author(s).) |
| Competing Interests: | Declarations. Conflict of interest: The authors have no competing interests to declare that are relevant to the content of this article. |
| References: | Nat Biomed Eng. 2021 Jun;5(6):555-570. (PMID: 33649564) Med Image Anal. 2019 Dec;58:101544. (PMID: 31466046) IEEE J Transl Eng Health Med. 2023 Feb 06;11:161-169. (PMID: 36816095) Health Inf Sci Syst. 2025 May 23;13(1):38. (PMID: 40416515) Diagnostics (Basel). 2024 Dec 23;14(24):. (PMID: 39767257) Med Image Anal. 2025 May;102:103504. (PMID: 40020420) BMC Bioinformatics. 2024 Mar 27;25(1):134. (PMID: 38539070) Biometrics. 1977 Mar;33(1):159-74. (PMID: 843571) Cell. 2020 Nov 25;183(5):1436-1456.e31. (PMID: 33212010) Clin Breast Cancer. 2025 Feb;25(2):e124-e132. (PMID: 39567339) Conf Comput Vis Pattern Recognit Workshops. 2021 Jun;2021:14318-14328. (PMID: 35047230) Breast Cancer Res. 2024 Aug 19;26(1):124. (PMID: 39160593) Mod Pathol. 2021 Apr;34(4):701-709. (PMID: 33077923) J Proteome Res. 2021 Apr 2;20(4):1902-1910. (PMID: 33560848) Commun Med (Lond). 2024 Dec 20;4(1):276. (PMID: 39706861) Cell Genom. 2021 Dec 8;1(3):. (PMID: 35465400) Health Inf Sci Syst. 2024 Nov 29;13(1):2. (PMID: 39619405) Comput Med Imaging Graph. 2025 Apr;121:102502. (PMID: 39919535) Cancer Res Treat. 2025 Jan;57(1):116-125. (PMID: 38938010) BMC Med Genomics. 2015 Aug 22;8:54. (PMID: 26297356) NPJ Precis Oncol. 2023 May 29;7(1):49. (PMID: 37248379) J Pathol Inform. 2024 Nov 17;16:100410. (PMID: 39720418) J Natl Cancer Cent. 2025 Feb 13;5(3):287-296. (PMID: 40693239) |
| Contributed Indexing: | Keywords: Breast cancer; Computational pathology; Domain shift; External validation; Foundation models |
| Substance Nomenclature: | 0 (Biomarkers, Tumor) EC 2.7.10.1 (Erb-b2 Receptor Tyrosine Kinases) |
| Entry Date(s): | Date Created: 20260511 Date Completed: 20260615 Latest Revision: 20260813 |
| Update Code: | 20260813 |
| PubMed Central ID: | PMC13269319 |
| DOI: | 10.1007/s11517-026-03590-4 |
| PMID: | 42113320 |
| Βάση Δεδομένων: | MEDLINE |
| ISSN: | 1741-0444 |
|---|---|
| DOI: | 10.1007/s11517-026-03590-4 |